# Figure replaced brittle robot code with neural networks

Markdown copy of https://realalpha.app/s/figure-replaced-brittle-robot-code-with-neural-networks/ on Real Alpha, the best minutes from long interviews with the people building AI. Cite and link the page, not this copy.

- Speaker: [Brett Adcock](https://realalpha.app/builders/brett-adcock/) · Brett Adcock is founder and CEO of Figure. · Founded Figure AI to build general-purpose humanoid robots; previously co-founded Archer Aviation and Vettery.
- Host: [Shawn Ryan](https://realalpha.app/people/shawn-ryan/) · Host, The Shawn Ryan Show
- Show: The Shawn Ryan Show
- Episode: [Brett Adcock - Shawn Ryan’s First Interview with a Robot | SRS #292](https://realalpha.app/v/99pOdGEGu6s/), Mar 30, 2026
- Moment: 1:12:00 to 1:16:28 of the episode (4 min), playing on https://realalpha.app/s/figure-replaced-brittle-robot-code-with-neural-networks/; the episode from there: https://realalpha.app/v/99pOdGEGu6s/?t=4320

## Quote

> “the robot that started the first day at the start of six months and **the robot that ended the shift** that day was the same”
> — [Brett Adcock](https://realalpha.app/builders/brett-adcock/) at [1:12:38](https://realalpha.app/v/99pOdGEGu6s/?t=4358) · The Shawn Ryan Show

## Takeaways

1. A Figure robot worked in a BMW factory for six months.
2. Code-based perception and control struggled with unfamiliar conditions.
3. Neural networks now power almost the entire robot stack.

## About this segment

After a six-month BMW factory trial, Brett Adcock says the robot hardware held up, but parts of the software struggled to scale. Figure moved nearly the whole stack to neural networks and now runs robots in continuous shifts.

From “Brett Adcock - Shawn Ryan’s First Interview with a Robot | SRS #292”, Mar 30, 2026. [Watch the full episode](https://realalpha.app/v/99pOdGEGu6s/)

## Transcript

Transcribed by machine (Whisper) from the episode’s audio, so a word may be off here and there.

[1:12:07](https://realalpha.app/v/99pOdGEGu6s/?t=4327) just like can we run robots every day you know are we gonna get like ass handed to us you know is it gonna be easy it's gonna be hard and uh i think it was in the middle i think we like um And we got the robot to a great spot where it was brand every day. It was great. I think the biggest learning lesson we took away is we had, I really cared about, can we do that? And can we like clone it times a thousand, times 10,000?

[1:12:29](https://realalpha.app/v/99pOdGEGu6s/?t=4349) Would we have any issues scaling? That was the part for me. Like, is it just like, you know, does it completely shit the bed and he's like, we need to rework our plan and go back to the office. Does it do it incredibly well? And you can just copy paste this thing to everywhere in the world. Like how does it, how did it work? And the biggest learning lesson we got was that the robots, the robot that started the first day at the start of six months and the robot that ended the shift that day was the same like even though we had multiple robots in operation every day we had like the same robot that did the start and the finish wow and it was cool like you know like it was same robot ended six months later and this was the thing where you know the worry was that humanoid robots couldn't last a month

[1:13:06](https://realalpha.app/v/99pOdGEGu6s/?t=4386) couldn't last a week in these kind of environments and uh you know wear and tear it's like it's running like a you know 40 degrees of freedom motors every single day can they operate really well. I think from a hardware perspective, it did an A plus job. I think from a software perspective, we did like a, I would say a B, B job, B plus. We, and that's mostly from like my perspective, like the architecture decisions I made to scale. We, about half the stack we had like traditional like code and heuristics in. So like the controller to walk was done by a C++ controller.

[1:13:41](https://realalpha.app/v/99pOdGEGu6s/?t=4421) it was done by code the walking you saw today we had back then was done in code okay the rest of this we had a bunch of other stuff in there done by neural nets and like some of the perception stacks some like you know some of how to move parts around everything else and you know this was a year and a half or so ago when we were first launching and um and i was like man this the the the biggest problems we're having is the coding parts get stuck the robot like either doesn't see like the right uh like doesn't like see something right on the part and misses the the object detector doesn't really understand what's going on the controller when it gets out of bounds of like what it's ever seen before like you have carpet in here now and it's like really squishy the robot's doing fine which is great but i think our old controller

[1:14:25](https://realalpha.app/v/99pOdGEGu6s/?t=4465) not do well it's like you have like really shaggy carpet here and it's like yeah and so like uh and it's like not very it's like it's like you know it's harder for a robot to walk around and um so that that that was like we had a really difficult time uh seeing that even though it did well every day seeing that scale to like lots of robots so we went back to the office this is about a year and a half ago and said we need to basically refactor everything into a neural network and one of the big like and i think we just announced helix two two or three months ago now I forget, like end of last year.

[1:14:56](https://realalpha.app/v/99pOdGEGu6s/?t=4496) And it's basically entirely down the stack, including the controllers and neural net now. There's like no code left really on the robot. Maybe some code in certain pieces, but mostly just almost all of the thing is like a neural net at this point. We removed the need for like almost over 100,000 lines of code when we launched Helix 2. And so what you saw today was just like a robot that we can put now back into, say, the factory and these places that will run only on a neural net.

[1:15:23](https://realalpha.app/v/99pOdGEGu6s/?t=4523) And I think we're running these robots right now. We're getting ready for deployment to customers and they're running incredibly well. We have robots running like basically now in 24-7 shifts without stopping, without any faults for like days and days. We just went like over, we just had like record time this past week on the robot running until we saw like a fault. Like almost, yeah, basically a whole week.

[1:15:49](https://realalpha.app/v/99pOdGEGu6s/?t=4549) Holy shit. Like they can run like four hours or so, five hours, and they need to charge. Another robot knows that, steps in, steps behind the robot, say, and gets ready for work. The robot then backs off. Another robot swaps and spot. Like in the next 10 seconds, it's doing work again. So we can run that now in 24-7 shifts where they're talking to each other, all autonomous. No humans or you can go to bed, whatever.

[1:16:13](https://realalpha.app/v/99pOdGEGu6s/?t=4573) And they're running that shifts all day and all night. And we do it across multiple use cases now at the office in 24-7 shifts. and it's just like we're running them hard what kind of stuff are they doing at your office we do